Nowcasting Economic Variables with Web Traffic Using VAR and Deep Learning

Overview

This repository provides code to predict economic variables using web traffic data.

Included Models:

• Causal Models: VAR
• Forecasting Models: ARIMA, Linear Model, VAR, LSTM, GRU

Getting Started

0. Prerequisites

• Python 3.12+ (Download Python)
• pip (comes with Python)

1. Clone the repository

git clone https://github.com/euro-kim/nowcast

2. Install dependencies

pip install -r requirements.txt

3. Prepare your data

• By default, the data contains monthly CPI, PPI, and employment data from South Korea (2010.01–2025.03).
• Google Trends data for the keyword '물가' and inflation in South Korea is also included.
• To use your own data, add it in CSV format to assets/data.csv.

4. Running the Main Script

The main entry point is run.py. You can use it to perform different activities with various models.

Basic syntax:

python run.py <activity> <model_name> [arguments]

Examples:

python run.py forecast gru --vars 'ppi,inflation' --seed 1
python run.py casual var --vars 'cpi,inflation'

Model Comparison Examples:

python run.py compare var --vars 'cpi,ppi'
python run.py compare_var_lstm_gru var --vars 'cpi,ppi'
python run.py compare_var_arima_gru var --vars 'cpi,ppi'
# Or using flags:
python run.py forecast gru --vars 'cpi,ppi' --compare_models
python run.py forecast gru --vars 'cpi,ppi' --compare_var_lstm_gru
python run.py forecast gru --vars 'cpi,ppi' --compare_var_arima_gru

---

Command-Line Arguments

Activities

Activity	Description
forecast	Forecasting (prediction)
casual	(removed)
compare	Compare ARIMA, AR, MA, VAR predictions
compare_var_lstm_gru	Compare VAR, LSTM, GRU predictions
compare_var_arima_gru	Compare VAR, ARIMA, GRU predictions

Models

model_name	Description
arima	ARIMA model
linear	Simple Linear Regression
var	VAR model
lstm	LSTM model
gru	GRU model

Arguments

Argument	Type	Default	Description
--seed	int	1	Random seed for reproducibility
--horizon	int	12	Number of time steps to forecast
--lag	int	12	Number of lagged observations (for VAR, LSTM, GRU)
--p	int	-1	AR order (AR, ARIMA, GARCH)
--d	int	-1	Differencing order (ARIMA)
--q	int	-1	MA order (MA, ARIMA, GARCH)
--maxlags	int	15	Maximum lags for VAR
--neurons	int	200	Number of neurons in RNN layers
--layers	int	1	Number of layers for RNN models
--batch_size	int	16	Batch size for RNN training
--epochs	int	100	Number of training epochs for RNN
--data_file	str	'assets/data.csv'	Path to the CSV data file
--vars	str	'cpi,ppi'	Comma-separated list of variables
--ic	str	'aic'	Information criterion for VAR (aic, bic, etc.)
--optimizer	str	'adam'	Optimizer for RNN models
--loss	str	'mean_squared_error'	Loss function for RNN models
--compare_models	flag		Compare ARIMA, AR, MA, VAR models in a single plot
--compare_var_lstm_gru	flag		Compare VAR, LSTM, GRU models in a single plot
--compare_var_arima_gru	flag		Compare VAR, ARIMA, GRU models in a single plot

---

Notes

• For model comparison, you can use either the activity argument (compare, compare_var_lstm_gru, compare_var_arima_gru) or the corresponding flag (--compare_models, --compare_var_lstm_gru, --compare_var_arima_gru).
• The --vars argument should be a comma-separated string of variable names present in your data file.
• All plots and results are saved in the results/ directory.

---